emacs-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Tree-sitter indentation for js-mode & cc-mode


From: Yuan Fu
Subject: Re: Tree-sitter indentation for js-mode & cc-mode
Date: Fri, 28 Oct 2022 12:43:31 -0700


> On Oct 28, 2022, at 2:10 AM, Theodor Thornhill <theo@thornhill.no> wrote:
> 
> Yuan Fu <casouri@gmail.com> writes:
> 
>>> 
>>>>> looking up way to much the root of the tree, but you know the internals
>>>>> here better than me.  Is this something we can optimize away? See the
>>>>> attached report at the bottom.
>>>> 
>>>> This is very strange, I need to look into it.
>>>> 
>>> 
>>> I'm happy to provide more info and profiling, as well as testing if need 
>>> be! 
>> 
>> I just tried running treesit-buffer-root-node and treesit-node-at
>> 10000 times in the end of buffer and they are pretty fast, so I don’t
>> know why the benchmark says 99% time is spent in
>> treesit-buffer-root-node. Could you share the benchmark code and test
>> file? Thanks!
>> 
>> Yuan
> 
> 
> Absolutely.  I ran the test again - see test file and new report in
> attachments.
> 
> You need to `M-x eval-buffer` in `treesit.el` to avoid the compiled
> functions to get better profile report, then in the testfile:
> 
> M-x profiler-start
> C-x h ;; (mark-whole-buffer)
> C-i   ;; (indent-for-tab-command)
> ;; --- waaaaait
> M-x profiler-stop
> M-x profiler-report
> 
> There's no test code for this, just running the commands sequentially
> and get the report :-)
> 
> Are we parsing the whole file over and over in treesit-buffer-root-node?
> Do we for some reason not hit the early return?
> 
> The js-file [0] is taken from [1] and duplicated and messed up
> indentation. Report [2] was messed up by dpaste it seems.  Gmail didn't
> want the files as attachments, so paste it is :-)
> 
> Hope this is useful!
> 
> Theo
> 

Ok, I’m fairly certain this is due to tree-sitter reparsing after we indenting 
each line: treesit-buffer-root-node asks for the root node of the parser, which 
triggers a reparse, because last indent modified the buffer. We are basically 
reparsing as many time as there are lines in the buffer.

Indenting a similarly sized buffer where all indent are good is much faster, 
because there is no reparse due to change to the buffer.

Tree-sitter indent should add an implementation for indent-for-region function 
which precomputes indent for each line and indent lines in batch. That ought to 
fix it. Added to TODO :-)

Yuan


reply via email to

[Prev in Thread] Current Thread [Next in Thread]